Topics:

• Data visualization seems to not be useful since we have an analytic solution for the displacements
• Mauricio shares this reference about surrogate modes to predict breast displacements field.
• Leo shows showed scripts to generate .csv results. Features [ $L_x$ , $E$ , $p$ ] targets [ $u_x$ , $u_y$ , $u_z$ ]
• Bruno contributes to the analytic solution of the uni-axial compression example.

Hice un primer informe pero hay muchas cosas que siguen siendo placeholder. En particular faltan

• Imágenes del problema (tanto uniaxial como cantilever) #43
• Incorporar las imágenes de loss
• Incorporar los datos de error y loss
• Lectura por alguien que no sea @leopoldoagorio
• Incorporar error para un ejemplo fuera del training set.

• Set the same number of samples for validation and test set. Moreover, check all test samples are inside the hypercube used for testing.
• Sample test not with a uniform distribution
• Compute the error versus the NN and ONSAS solution for an outlier (most flexible case).

• ONSAS.m script
• Create dataset
• Surrogate model train and test
  

This example should include:

• Fixed transversal loads Py, Pz
• Vary compression $p$

Useful reads 📖:

• Notebook

• One trial form HeatTransferPINN

In the training loop we can use:

# Training
for epoch in range(EPOCHS):
    gen.train()
    dis.train()

    for i, data in enumerate(training_loader):

In order to avoid this model call, we can define an inner call method to convert mlp object into a functor like this:

    def __call__(self, x):
        return self.forward(x)

Instead of:

mlp.Net(X)

We can use:

mlp(X)

This function must compute the target values with the analytical solution for u_y and u_z given the input features

We had discussed this but i don't remember if we reached a veredict

Not what was expected yet.
Are we sure we are passing p and not -p in either analytic solution or model evaluation?
Please read and check

Que quede claro que es un único punto para el cual se trata de aprender el desplazamiento en funcion de los parametros

• Abstract
• ANNs methodology
• Integration with ONSAS.jl
• [ ]

| xi -yi | / | yi |

Define areas

Add papers

Inlcude them into the .tex

• Intro (Refs, describiendo (1 parrafo minimo)
• Herramientas utilizadas (1 parrafo minimo)
• Resutlados numéricos
• Conclusiones

We have exact solution
Uniaxial Compression
can test in grid against that solution solving numerically the non-linear equation

The github users are:

Jorge Pérez Zerpa: jorgepz
Bruno Bazzano: bbazzano

File:
https://amcaonline.org.ar/ojs/index.php/mc/article/download/6605/6668

Did we added an activation function on the first layer output ?

• Plot loss vs epchs
• Dev set error

As opened in #15 discussion

We had discussed this but i don't remember if we reached a veredict

https://towardsdatascience.com/what-is-k-fold-cross-validation-5a7bb241d82f

Use tqdm for progress bar

Not able to produce menaningful comparisons. Must check in detail.

• Try to eval MLP and analytic in meaningful points
• Debug dataset generation
• Check math

See loss plot for 100 epochs (train/val/analytic) to understand the error